Python Pandas：矢量化操作错误？

Question

Date series looks something like this.

In [89]:
db.close[:5]

Out[89]:
datetime
2012-06-28 23:58:00    1.243925
2012-06-28 23:59:00    1.244125
2012-06-29 00:00:00    1.244065
2012-06-29 00:01:00    1.243875
2012-06-29 00:02:00    1.243865
Name: close

I would like to subtract previous element from each element.

In [93]:
db.close[1:5] - db.close[:4]

Out[93]:
datetime
2012-06-28 23:58:00   NaN
2012-06-28 23:59:00     0
2012-06-29 00:00:00     0
2012-06-29 00:01:00     0
2012-06-29 00:02:00   NaN
Name: close

Arrays was subtract without offset.

But when I compare the array elements...

n [94]:

db.close[1:5] == db.close[:4]
Out[94]:
datetime
2012-06-28 23:59:00    False
2012-06-29 00:00:00    False
2012-06-29 00:01:00    False
2012-06-29 00:02:00    False
Name: close

Answer 1

This is actually deliberate. Arithmetic operations do data alignment, but comparisons do not. I considered changing it in the past but found that it caused too many problems (especially when passing Series to functions expecting NumPy arrays, as an example, numpy.diff ).

EDIT: to get alignment, you can do the alignment by hand:

In [10]: numpy.equal(*a.align(b))
Out[10]: 
2000-01-03    False
2000-01-04     True
2000-01-05     True
2000-01-06     True
2000-01-07     True
2000-01-10    False
Freq: B

Answer 2

I found the answer to my own question. It may be useful to someone.

In [38]:
db.close.shift(periods=1).head() - db.close.head()

Out[38]:
datetime
2012-06-28 23:58:00        NaN
2012-06-28 23:59:00   -0.00020
2012-06-29 00:00:00    0.00006
2012-06-29 00:01:00    0.00019
2012-06-29 00:02:00    0.00001
Freq: T, Name: close

Unfortunately, it is 2-3x slower than normal arithmetic operations.

In [40]:

%timeit db.close.shift(periods=1) - db.close
1000 loops, best of 3: 1.64 ms per loop

%timeit db.close - db.open
1000 loops, best of 3: 700 us per loop